tacit/nanochat
1
0
Fork 0
mirror of https://github.com/karpathy/nanochat.git synced 2025-12-25 21:52:26 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix gb200 tflops

Browse Source
This commit is contained in:
Qubitium 2025-10-22 02:03:23 +00:00
parent 4e6f5eb8b9
commit 32017e831a
1 changed files with 8 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
nanochat/mfu.py
View File

@ -4,7 +4,6 @@ import torch
def get_promised_flops_per_gpu(): def get_promised_flops_per_gpu():
""" """
Return best-effort dense BF16 Tensor/Matrix peak FLOPs for the active GPU. Return best-effort dense BF16 Tensor/Matrix peak FLOPs for the active GPU.
Returns: Returns:
tuple[str, float, bool]: (device_name, flops_per_gpu, is_estimated) tuple[str, float, bool]: (device_name, flops_per_gpu, is_estimated)
device_name is the CUDA-reported name for the active device. device_name is the CUDA-reported name for the active device.
@ -23,29 +22,29 @@ def get_promised_flops_per_gpu():
return any(k in name for k in keywords) return any(k in name for k in keywords)
# --- NVIDIA Blackwell --- # --- NVIDIA Blackwell ---
if has("gb200", "grace blackwell", "b200"): if has("gb200", "grace blackwell"):
return result(2.25e15, False) # B200 dense BF16 ≈ 2.25 PFLOPS. (4.5 PFLOPS sparse) return result(2.5e15, False) # GB200 dense BF16 ≈ 2.5 PFLOPS (5.0 PFLOPS sparse)
if has("b200"):
return result(2.25e15, False) # B200 dense BF16 ≈ 2.25 PFLOPS (4.5 PFLOPS sparse)
if has("b100"): if has("b100"):
return result(1.8e15, False) # B100 dense BF16 ≈ 1.8 PFLOPS. (3.5 PFLOPS sparse) return result(1.8e15, False) # B100 dense BF16 ≈ 1.8 PFLOPS (3.5 PFLOPS sparse)
# --- NVIDIA Hopper (H100/H200/H800) --- # --- NVIDIA Hopper (H100/H200/H800) ---
if has("h200"): if has("h200"):
# SXM ~= H100 SXM for compute; NVL/PCIe is lower clock
if has("nvl", "pcie"): if has("nvl", "pcie"):
return result(836e12, False) # H200 NVL/PCIe dense BF16 ≈ 836 TFLOPS return result(836e12, False) # H200 NVL/PCIe dense BF16 ≈ 836 TFLOPS
return result(989e12, False) # H200 SXM dense BF16 ≈ 989 TFLOPS return result(989e12, False) # H200 SXM dense BF16 ≈ 989 TFLOPS
if has("h100"): if has("h100"):
if has("nvl"): # H100 NVL (per GPU) if has("nvl"):
return result(835e12, False) # 1671 TFLOPS (sparse) ⇒ ~835 dense return result(835e12, False) # H100 NVL dense BF16 ≈ 835 TFLOPS
if has("pcie"): if has("pcie"):
return result(756e12, False) # H100 PCIe dense BF16 ≈ 756 TFLOPS return result(756e12, False) # H100 PCIe dense BF16 ≈ 756 TFLOPS
return result(989e12, False) # H100 SXM dense BF16 ≈ 989 TFLOPS return result(989e12, False) # H100 SXM dense BF16 ≈ 989 TFLOPS
if has("h800"): if has("h800"):
# China-optimized Hopper; NVLink configs often quoted at H100-SXM-like numbers
if has("nvl"): if has("nvl"):
return result(989e12, False) # H800 NVLink dense BF16 ≈ 989 TFLOPS (vendor configs) return result(989e12, False) # H800 NVLink dense BF16 ≈ 989 TFLOPS
return result(756e12, False) # H800 PCIe dense BF16 ≈ 756 TFLOPS return result(756e12, False) # H800 PCIe dense BF16 ≈ 756 TFLOPS
# --- NVIDIA Ampere data center / export variants --- # --- NVIDIA Ampere data center / export variants ---